A Deep Learning Method For Image Super Resolution Based On Geometric Similarity

With the development and progress of science and technology,image as the main carrier of visual information has been paid more and more attention.Under the premise of not changing the existing image acquisition system,the image super-resolution reconstruction technology,which uses the digital image processing technology to enhance the spatial resolution,has attracted more and more attention from scholars at home and abroad.Image super-resolution reconstruction algorithm is to restore a high resolution image with high quality and rich details from a known low resolution image,Its main task is how to reconstruct high-frequency detail information that is lost during the observation process.High-resolution image are rich in content and can provide more useful information for other computer vision problems such as target recognition and detection.It has been widely used in surveillance,public security,remote sensing,medical imaging,etc.Based on the learning method,the mapping function relationship between the low-resolution image and the high-resolution image is obtained from the large number of image samples by the machine learning algorithm,and then the mapping function relation is used to get the high frequency detail information,which is needed to reconstruct the image.This method has achieved excellent results and has gradually become a hotspot in this field.Therefore,we focuses on the study of image super-resolution image reconstruction algorithm based on learning in this paper.(1)Aiming at the problem that the existing convolution neural network for super-resolution algorithm can not further supplement more detailed information,therefore,we presents an image super-resolution algorithm combining image multi-scale self-similarity and depth convolution neural network in this paper,and use the large number of repetitive regions existing in the image to supplement the reconstruction area,thus the high-resolution image reconstructed has rich details and good visibility.(2)Based on the multi-scale self-similarity of images,we look for similar image blocks in the whole image that can be used to reconstruct the details for each image block to be reconstructed.These similar image blocks contain not only similar image blocks with translation transform,but also those Similar image blocks with scale,rotation transform.Then,these similar image blocks are effectively merged together.In this paper,the spatial transform network layer is used to transform these similar image blocks,thus the similar information in the similar image blocks is fully utilized.(3)In this paper,we use the deconvolution layer instead of the bi-cubic interpolation kernel to amplify the image and to estimate the initial of the high-resolution image.And the multi convolution layer are used to extract the details of the image,and then we use the pyramid model of convolution network to gradually predict the high-frequency image details,thus the final high-resolution image is achieved from the coarse to fine progressive mode.Extensive experiment results demonstrate the effective and efficiency of the proposed network structure.In subjective evaluation,the reconstructed image is rich in detail and has good visual effect.